adamo1139/LLaMa-30B-AEZAKMI-v2-LoRA

Overview

Llama 30B/33B fine-tuned on AEZAKMI v2 dataset. I will be releasing only adapter files for now, maybe weights will come later. It works nicely, it's answers seem to be shorter than those of Yi-34B but fully within reason.

Prompt template is chatml, same as my other models.

<|im_start|>system
A chat with uncensored assistant.<|im_end|>
<|im_start|>user
{prompt}<|im_end|>
<|im_start|>assistant

You can find answers to some test prompts here https://huggingface.co/datasets/adamo1139/misc/blob/main/benchmarks/llama-30b-aezakmi-v2/llama-30b-aezakmi-v2.txt

Fine-tuning details

base_model: ./llama-30b
base_model_config: ./llama-30b
model_type: LlamaForCausalLM
tokenizer_type: LlamaTokenizer
is_mistral_derived_model: false
is_llama_derived_model: true

load_in_8bit: false
load_in_4bit: true

torch_dtype: bf16
strict: false
datasets:
  - path: /run/media/..../axolotl/datasets/aezakmi_v2/aezakmi_v2_draft2.jsonl
    type: alpaca_w_system2.load_open_orca_chatml
    conversation: chatml
dataset_prepared_path: last_run_prepared
val_set_size: 0.01
adapter: qlora
lora_model_dir:
sequence_len: 1000
sample_packing: true
lora_r: 16
lora_alpha: 32
lora_dropout: 0.05
lora_target_modules:
  - q_proj
  - v_proj
  - k_proj
  - o_proj
  - gate_proj
  - down_proj
  - up_proj
lora_target_linear: true
lora_fan_in_fan_out:
wandb_project:
wandb_watch:
wandb_run_id:
wandb_log_model:
output_dir: ./qlora-llama-30b-aezakmi-v2-run2
pad_to_sequence_len: true
lora_modules_to_save:
 - embed_tokens
 - lm_head
micro_batch_size: 1
gradient_accumulation_steps: 1
num_epochs: 2
optimizer: adamw_bnb_8bit
torchdistx_path:
lr_scheduler: cosine
learning_rate: 0.00001
train_on_inputs: false
group_by_length: false
bf16: true
fp16: false
tf32: false
bfloat16: true
flash_optimum: false
gradient_checkpointing: true
early_stopping_patience:
save_safetensors: true
local_rank:
logging_steps: 1
xformers_attention:
flash_attention: true
deepspeed:
seed: 42
warmup_steps: 100
eval_steps: 5000000
save_steps: 1000
save_total_limit: 10
eval_table_size: 
eval_table_max_new_tokens:
debug:
weight_decay:
fsdp:
fsdp_config:
special_tokens:
  bos_token: "<s>"
  eos_token: "</s>"
  unk_token: "<unk>"
tokens:
 - "<|im_end|>"
 - "<|im_start|>"